Deep Learning for Automated Defect Detection in Industrial Manufacturing

The automation of defect detection in industrial
manufacturing is essential for quality assurance, as well as
reducing manual inspection costs. Traditional inspection
approaches often have challenges in scalability, accuracy, and
adapting to complex defect types. To overcome these
challenges, this study proposes a deep learning framework
called M2U-InspectNet which takes advantage of multi-scale
vision transformers, self-supervised contrastive pretraining,
and an uncertainty-aware defect localization module. The
model is trained and evaluated on the publicly available
MVTec AD dataset, which includes a variety of industrial
textures and objects. Experimental results show that M2UInspectNet achieved a 94.8% accuracy with 91.7% mean
Average Precision (mAP) and retained a real-time inference
speed of 52 Frames per Second (FPS) on edge devices,
outperforming baseline existing CNN and object detection
methods. Notably, M2U-InspectNet also performed well under
limited data robustness testing, indicating the potential to be used in real-world industrial contexts. The findings of this research suggest the utility of transformer-based architectures as well as self-supervised learning for enhancing visual inspection systems to make them smarter, more scalable, and more reliable for manufacturing quality control in the future.